diff --git a/internal/cli/coverage_test.go b/internal/cli/coverage_test.go
index 152beaf..107d3a1 100644
--- a/internal/cli/coverage_test.go
+++ b/internal/cli/coverage_test.go
@@ -83,6 +83,24 @@ func TestExitError_Methods(t *testing.T) {
 	}
 }
 
+// TestRunDatasetRm_InvalidTableExitsTwo: rm validates the table name
+// before any cluster work, so an unsafe name is exit-code-2 territory
+// and never reaches kubeconfig/cluster resolution.
+func TestRunDatasetRm_InvalidTableExitsTwo(t *testing.T) {
+	var buf bytes.Buffer
+	err := runDatasetRm(context.Background(), runDatasetRmArgs{
+		Table:   "../bad",
+		Printer: ui.New(&buf, ui.WithColor(false)),
+	})
+	var ee *exitError
+	if !errors.As(err, &ee) {
+		t.Fatalf("error is not an *exitError: %v", err)
+	}
+	if ee.Code() != 2 {
+		t.Errorf("exit code = %d, want 2 (invalid table name)", ee.Code())
+	}
+}
+
 // TestRunClusterInfo_BadKubeconfigExitsThree: an unreadable/invalid
 // kubeconfig is exit-code-3 territory (the kubeconfig/local-input
 // bucket), surfaced before any cluster work. Covers the Load-error
diff --git a/internal/cli/dataset.go b/internal/cli/dataset.go
index 6a8b098..55561b3 100644
--- a/internal/cli/dataset.go
+++ b/internal/cli/dataset.go
@@ -22,8 +22,9 @@ import (
 // verb is `push`, completed in Phase 3 (tracebloc/client#151) across
 // PR-a (pre-flight: spec synth, validation, layout walk, cluster
 // discovery) and PR-b (this one: ephemeral stage Pod + tar-over-
-// exec stream + progress bar + SIGINT-safe cleanup). Future verbs
-// (`dataset list`, `dataset rm`) hang off this parent in v0.2.
+// exec stream + progress bar + SIGINT-safe cleanup). `dataset rm`
+// (#30) removes a pushed dataset's in-cluster artifacts; `dataset
+// list` hangs off this parent later.
 func newDatasetCmd() *cobra.Command {
 	cmd := &cobra.Command{
 		Use:   "dataset",
@@ -39,6 +40,7 @@ ingestor Job to completion (streaming logs + the final summary).
 before the first push.`,
 	}
 	cmd.AddCommand(newDatasetPushCmd())
+	cmd.AddCommand(newDatasetRmCmd())
 	return cmd
 }
 
diff --git a/internal/cli/dataset_rm.go b/internal/cli/dataset_rm.go
new file mode 100644
index 0000000..5d78c01
--- /dev/null
+++ b/internal/cli/dataset_rm.go
@@ -0,0 +1,186 @@
+package cli
+
+import (
+	"context"
+	"errors"
+	"fmt"
+
+	"github.com/spf13/cobra"
+
+	"github.com/tracebloc/cli/internal/cluster"
+	"github.com/tracebloc/cli/internal/push"
+	"github.com/tracebloc/cli/internal/ui"
+)
+
+// runDatasetRmArgs is the resolved input to runDatasetRm — same shape
+// convention as runDatasetPushArgs, so the command's RunE stays a thin
+// flag-to-struct adapter and the logic is unit-testable.
+type runDatasetRmArgs struct {
+	Table      string
+	Kubeconfig string
+	Context    string
+	Namespace  string
+	DryRun     bool
+	Yes        bool
+	Printer    *ui.Printer
+	Prompter   prompter // nil off a TTY or when --yes is set
+}
+
+// newDatasetRmCmd implements `tracebloc dataset rm <table>` — the
+// in-cluster teardown of a previously-pushed dataset. See
+// internal/push.Teardown for the mechanism and the design note on the
+// approach (CLI-direct vs a server-side delete endpoint).
+func newDatasetRmCmd() *cobra.Command {
+	var (
+		kubeconfigPath  string
+		contextOverride string
+		nsOverride      string
+		dryRun          bool
+		yes             bool
+	)
+
+	cmd := &cobra.Command{
+		Use:   "rm <table>",
+		Short: "Delete a pushed dataset's in-cluster artifacts (table + PVC files)",
+		Long: `Removes the in-cluster artifacts a previous ` + "`dataset push`" + ` created
+for a table: the MySQL table in ` + push.IngestionDatabase + ` and the dataset's
+directories on the shared PVC. Destructive and not undoable.
+
+NOTE: the central tracebloc backend catalog entry is NOT removed — the
+CLI has no direct line to that backend. Full cleanup of a successfully
+ingested dataset needs the server-side delete path (tracebloc/cli#39).
+
+Exit codes:
+  0  artifacts removed (or --dry-run, or the user declined)
+  2  invalid table name
+  3  kubeconfig error, or refused (no confirmation off a terminal)
+  4  cluster reachable but parent release / shared PVC missing
+  7  teardown failed mid-flight (table drop or PVC rm errored)`,
+		Args: cobra.ExactArgs(1),
+		RunE: func(cmd *cobra.Command, args []string) error {
+			// Confirm interactively on a TTY unless --yes was passed.
+			var pr prompter
+			if !yes && isInteractiveTTY() {
+				pr = surveyPrompter{}
+			}
+			return runDatasetRm(cmd.Context(), runDatasetRmArgs{
+				Table:      args[0],
+				Kubeconfig: kubeconfigPath,
+				Context:    contextOverride,
+				Namespace:  nsOverride,
+				DryRun:     dryRun,
+				Yes:        yes,
+				Printer:    printerFor(cmd),
+				Prompter:   pr,
+			})
+		},
+	}
+
+	cmd.Flags().StringVar(&kubeconfigPath, "kubeconfig", "",
+		"path to the kubeconfig file (default: $KUBECONFIG, then ~/.kube/config)")
+	cmd.Flags().StringVar(&contextOverride, "context", "",
+		"name of the kubeconfig context to use (default: kubeconfig's current-context)")
+	cmd.Flags().StringVarP(&nsOverride, "namespace", "n", "",
+		"namespace where the parent tracebloc/client release is installed")
+	cmd.Flags().BoolVar(&dryRun, "dry-run", false,
+		"show what would be deleted without deleting anything")
+	cmd.Flags().BoolVarP(&yes, "yes", "y", false,
+		"skip the confirmation prompt (required when not on a terminal)")
+
+	return cmd
+}
+
+// runDatasetRm discovers the cluster, shows the teardown plan, confirms,
+// then removes the in-cluster artifacts. The flow mirrors runDatasetPush
+// (validate → discover → plan/pre-flight → act) so the two commands feel
+// like siblings.
+func runDatasetRm(ctx context.Context, a runDatasetRmArgs) error {
+	p := a.Printer
+	p.Banner("tracebloc", "delete a pushed dataset")
+
+	// 1. Validate the name before we build any PVC path from it
+	//    (push.PlanTeardown panics on an unsafe name by design).
+	if err := push.ValidateTableName(a.Table); err != nil {
+		return &exitError{code: 2, err: fmt.Errorf("invalid table name %q: %w", a.Table, err)}
+	}
+
+	// 2. Resolve cluster + clientset (kubeconfig errors = exit 3).
+	resolved, err := cluster.Load(cluster.KubeconfigOptions{
+		Path:      a.Kubeconfig,
+		Context:   a.Context,
+		Namespace: a.Namespace,
+	})
+	if err != nil {
+		return &exitError{code: 3, err: fmt.Errorf("loading kubeconfig: %w", err)}
+	}
+	cs, err := cluster.NewClientset(resolved)
+	if err != nil {
+		return &exitError{code: 3, err: err}
+	}
+
+	// 3. Confirm the parent release + shared PVC exist (exit 4 if not) —
+	//    both "is this the right cluster?" context and a guard against
+	//    running teardown against a cluster with no tracebloc install.
+	release, err := cluster.DiscoverParentRelease(ctx, cs, resolved.Namespace)
+	if err != nil {
+		return &exitError{code: 4, err: err}
+	}
+	pvc, err := cluster.DiscoverSharedPVC(ctx, cs, resolved.Namespace)
+	if err != nil {
+		return &exitError{code: 4, err: err}
+	}
+
+	// 4. Show exactly what will be deleted — the customer's last look
+	//    before destructive, unrecoverable work.
+	plan := push.PlanTeardown(a.Table)
+
+	p.Section("Target")
+	p.Field("context", resolved.Context)
+	p.Field("namespace", resolved.Namespace)
+	p.Field("release", release.ReleaseName)
+	p.Field("shared PVC", pvc.ClaimName)
+
+	p.Section("Will delete")
+	p.Field("mysql table", plan.Database+"."+plan.Table)
+	for _, path := range plan.PVCPaths {
+		p.Field("pvc path", path)
+	}
+	p.Warnf("Destructive and cannot be undone. The central backend catalog entry is NOT removed (tracebloc/cli#39).")
+
+	// 5. Dry-run stop.
+	if a.DryRun {
+		p.Successf("Dry-run — nothing was deleted.")
+		return nil
+	}
+
+	// 6. Confirm. --yes skips; off a TTY without --yes we refuse rather
+	//    than delete unprompted.
+	if !a.Yes {
+		if a.Prompter == nil {
+			return &exitError{code: 3, err: errors.New(
+				"refusing to delete without confirmation: pass --yes or run on a terminal")}
+		}
+		ok, err := a.Prompter.Confirm(fmt.Sprintf("Delete %q and its files?", a.Table), false)
+		if err != nil {
+			if errors.Is(err, errInteractiveCancelled) {
+				p.Infof("Cancelled — nothing was deleted.")
+				return nil
+			}
+			return &exitError{code: 3, err: err}
+		}
+		if !ok {
+			p.Infof("Cancelled — nothing was deleted.")
+			return nil
+		}
+	}
+
+	// 7. Execute the in-cluster teardown.
+	p.Infof("Removing in-cluster artifacts…")
+	res, err := push.Teardown(ctx, cs, resolved.RestConfig, resolved.Namespace, plan)
+	if err != nil {
+		return &exitError{code: 7, err: fmt.Errorf("teardown failed: %w", err)}
+	}
+
+	p.Successf("Deleted %s.%s and %d PVC path(s).", plan.Database, plan.Table, len(res.RemovedPaths))
+	return nil
+}
diff --git a/internal/push/teardown.go b/internal/push/teardown.go
new file mode 100644
index 0000000..6f3a9c9
--- /dev/null
+++ b/internal/push/teardown.go
@@ -0,0 +1,125 @@
+package push
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"strings"
+
+	corev1 "k8s.io/api/core/v1"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/client-go/kubernetes"
+	"k8s.io/client-go/rest"
+)
+
+// IngestionDatabase is the MySQL schema jobs-manager ingests tables
+// into (data-ingestors' target DB). Centralized here so the push path
+// and the teardown path agree on where a table lives.
+const IngestionDatabase = "training_test_datasets"
+
+// TeardownPlan enumerates the in-cluster artifacts `dataset rm` removes
+// for a pushed table: the MySQL table and the dataset's directories on
+// the shared PVC.
+//
+// It deliberately does NOT include the central tracebloc backend
+// catalog entry: the CLI has no direct line to that backend (only the
+// in-cluster ingestor does, with its own creds), so removing it is the
+// cross-repo follow-up (tracebloc/cli#39). A successfully-ingested
+// dataset torn down this way leaves a stale catalog entry until #39
+// lands.
+type TeardownPlan struct {
+	Database string   // MySQL schema (IngestionDatabase)
+	Table    string   // table name — MUST have passed ValidateTableName
+	PVCPaths []string // absolute dirs on the shared PVC to rm -rf
+}
+
+// PlanTeardown computes the teardown plan for a table. It calls
+// FinalDestPrefix/StagedPrefix, which panic on an unsafe name — every
+// caller validates with ValidateTableName first (see cli.runDatasetRm).
+func PlanTeardown(table string) TeardownPlan {
+	return TeardownPlan{
+		Database: IngestionDatabase,
+		Table:    table,
+		PVCPaths: []string{FinalDestPrefix(table), StagedPrefix(table)},
+	}
+}
+
+// TeardownResult reports what Teardown actually removed.
+type TeardownResult struct {
+	DroppedTable bool
+	RemovedPaths []string
+}
+
+// Teardown performs the in-cluster teardown described by plan, mirroring
+// the manual kubectl-exec cleanup:
+//
+//   - DROP the MySQL table by exec-ing `mysql` inside the mysql pod,
+//     referencing the pod's own $MYSQL_ROOT_PASSWORD — so no database
+//     credential ever transits the CLI.
+//   - rm -rf the PVC dirs by exec-ing inside the jobs-manager pod,
+//     which mounts the shared PVC at SharedRoot.
+//
+// DESIGN NOTE (under review): this exec-into-existing-pods approach is
+// the "CLI-direct teardown". The alternative under discussion is a
+// server-side jobs-manager delete endpoint that could also remove the
+// backend catalog entry (#39) in one place. It assumes (a) a pod whose
+// name contains "mysql" exposes $MYSQL_ROOT_PASSWORD, and (b) the
+// jobs-manager pod mounts the shared PVC at SharedRoot — both true for
+// the current parent chart, but worth confirming before this ships.
+func Teardown(ctx context.Context, cs kubernetes.Interface, cfg *rest.Config, namespace string, plan TeardownPlan) (TeardownResult, error) {
+	var res TeardownResult
+	exec := &SPDYExecutor{Config: cfg, Client: cs}
+
+	// 1. DROP the table — mysql pod, localhost, its own root password.
+	mysqlPod, mysqlContainer, err := findRunningPod(ctx, cs, namespace, "mysql")
+	if err != nil {
+		return res, fmt.Errorf("locating mysql pod: %w", err)
+	}
+	sql := fmt.Sprintf("DROP TABLE IF EXISTS `%s`.`%s`", plan.Database, plan.Table)
+	script := fmt.Sprintf(`mysql -uroot -p"$MYSQL_ROOT_PASSWORD" -e '%s'`, sql)
+	var stderr bytes.Buffer
+	if err := exec.Exec(ctx, namespace, mysqlPod, mysqlContainer,
+		[]string{"sh", "-c", script}, nil, nil, &stderr); err != nil {
+		return res, fmt.Errorf("dropping %s.%s: %w%s", plan.Database, plan.Table, err, stderrSuffix(&stderr))
+	}
+	res.DroppedTable = true
+
+	// 2. rm the PVC dirs — jobs-manager pod mounts the shared PVC.
+	jmPod, jmContainer, err := findRunningPod(ctx, cs, namespace, "jobs-manager")
+	if err != nil {
+		return res, fmt.Errorf("locating jobs-manager pod: %w", err)
+	}
+	stderr.Reset()
+	rmCmd := append([]string{"rm", "-rf"}, plan.PVCPaths...)
+	if err := exec.Exec(ctx, namespace, jmPod, jmContainer, rmCmd, nil, nil, &stderr); err != nil {
+		return res, fmt.Errorf("removing PVC paths: %w%s", err, stderrSuffix(&stderr))
+	}
+	res.RemovedPaths = plan.PVCPaths
+	return res, nil
+}
+
+// findRunningPod returns the name + first-container name of the first
+// Running pod in namespace whose name contains substr.
+func findRunningPod(ctx context.Context, cs kubernetes.Interface, namespace, substr string) (podName, container string, err error) {
+	pods, err := cs.CoreV1().Pods(namespace).List(ctx, metav1.ListOptions{})
+	if err != nil {
+		return "", "", err
+	}
+	for i := range pods.Items {
+		p := &pods.Items[i]
+		if p.Status.Phase == corev1.PodRunning && strings.Contains(p.Name, substr) && len(p.Spec.Containers) > 0 {
+			return p.Name, p.Spec.Containers[0].Name, nil
+		}
+	}
+	return "", "", fmt.Errorf("no Running pod with name containing %q in namespace %q", substr, namespace)
+}
+
+// stderrSuffix renders captured stderr as a parenthetical for error
+// messages, or "" when the remote command was quiet.
+func stderrSuffix(b *bytes.Buffer) string {
+	s := strings.TrimSpace(b.String())
+	if s == "" {
+		return ""
+	}
+	return " (" + s + ")"
+}
diff --git a/internal/push/teardown_test.go b/internal/push/teardown_test.go
new file mode 100644
index 0000000..a3e9444
--- /dev/null
+++ b/internal/push/teardown_test.go
@@ -0,0 +1,30 @@
+package push
+
+import "testing"
+
+// TestPlanTeardown pins the artifact set `dataset rm` targets: the
+// MySQL table in IngestionDatabase + both PVC dirs (final dest +
+// staging), in that order.
+func TestPlanTeardown(t *testing.T) {
+	plan := PlanTeardown("reg_train")
+
+	if plan.Database != IngestionDatabase {
+		t.Errorf("Database = %q, want %q", plan.Database, IngestionDatabase)
+	}
+	if plan.Table != "reg_train" {
+		t.Errorf("Table = %q, want reg_train", plan.Table)
+	}
+
+	want := []string{
+		"/data/shared/reg_train",
+		"/data/shared/.tracebloc-staging/reg_train",
+	}
+	if len(plan.PVCPaths) != len(want) {
+		t.Fatalf("PVCPaths = %v, want %v", plan.PVCPaths, want)
+	}
+	for i := range want {
+		if plan.PVCPaths[i] != want[i] {
+			t.Errorf("PVCPaths[%d] = %q, want %q", i, plan.PVCPaths[i], want[i])
+		}
+	}
+}